RafiqPrinciples
These are the commitments every Rafiq product is held to. A change that breaks one of them is rejected regardless of how useful it is. Changing a principle itself requires a public RFC.
Local by default
Inference, storage, search and memory work fully offline on a laptop or a home server. No account is needed to start. Cloud is an option the user turns on.
Consent per data flow
Every outbound connection is a named, revocable grant visible in one place. Nothing leaves the device without a grant the user can see and withdraw.
No telemetry unless opted in
Opt-in telemetry is anonymous, inspectable before it is sent, and documented field by field.
Encrypted at rest and end to end
Keys stay with the user. The reference sync relay never holds plaintext, and users can run their own relay.
Data portability
Markdown, JSON, SQLite and standard calendar and contact formats. One export produces everything a user has put in or Rafiq has learned.
Open weights preferred, cloud allowed
The gateway treats every model the same. The user always sees which model answered and where it ran.
Every language is first class
Interface, OCR, transcription, embeddings and evaluation sets are built for more than one language and writing direction from the first release.
Security is a product feature
Published threat model, signed and reproducible builds, responsible disclosure, external audit before 1.0, and a user-readable audit log.
Readable, editable memory
Memory is a document the user owns, not a hidden embedding store.
No dark patterns
No nagging to enable cloud, no degraded local mode to push upgrades, no engagement optimisation. Rafiq should be the companion you forget is running.